armature-harness 0.2.3

Harness engineering framework for AI coding agents -- the invisible skeleton that shapes agent output

Armature

The invisible skeleton that gives shape to what AI coding agents produce.

Armature is a harness engineering framework for AI coding agents. It wraps agents (Claude Code, Cursor, Copilot, Windsurf, Aider) in automated guides, sensors, budget controls, architectural enforcement, garbage collection, and self-healing pipelines.

System Architecture

Usage Flow

Quick Start

pip install armature-harness

# Initialize in your project
cd your-project/
armature init

# Run quality checks
armature check

# Wire into Claude Code
armature hooks --claude-code

# Self-heal lint violations
armature heal --failures lint

The 6 Pillars

Pillar	What	CLI
Budget	Token/cost tracking, multi-provider routing, semantic cache, auto-calibration	armature budget
Quality	8 weighted checks (lint, type, test, security, complexity, deps, docstring, ratio)	armature check
Context	CLAUDE.md/AGENTS.md generation, progressive disclosure, cross-session memory	armature hooks
Architecture	Layer boundary enforcement, class conformance, schema sync	armature check
GC	Architecture drift, stale docs, dead code, budget audit	armature gc
Self-Heal	Auto-fix lint, report type/test errors, circuit breaker escalation	armature heal

Quality Checks (v0.2.1)

8 weighted checks with configurable gates:

Check	Tool	Weight	Type
Lint	ruff	25	External
Type-check	mypy	25	External
Tests	pytest	20	External
Security	bandit	20	External
Complexity	radon	15	External
Dependency audit	pip-audit	15	External
Docstring coverage	AST analysis	10	Internal
Test-to-code ratio	LOC analysis	10	Internal

Scoring: weighted_score = sum(score * weight) / sum(weight)

Gates: Draft (70%) | Review Ready (85%) | Merge Ready (95%)

Configuration

Everything is configured in armature.yaml:

project:
  name: "my-project"
  language: python
  src_dir: "src/"

quality:
  enabled: true
  checks:
    lint: { tool: ruff, weight: 25 }
    type_check: { tool: mypy, weight: 25 }
    test: { tool: pytest, weight: 20, coverage_min: 85 }
    security: { tool: bandit, weight: 20 }
    complexity: { kind: internal, weight: 15, threshold: 10.0 }
    dependency_audit: { tool: pip-audit, weight: 15 }
    docstring: { kind: internal, weight: 10 }
    test_ratio: { kind: internal, weight: 10, threshold: 0.5 }
  post_write:
    enabled: true  # shift-left: check on every file write

budget:
  enabled: true
  providers:
    strategy: cost_optimized
    enabled_models: [claude-sonnet, claude-haiku, claude-opus]
  cache: { enabled: true }
  calibration: { enabled: true, auto_calibrate: true }

architecture:
  enabled: true
  layers:
    - { name: models, dirs: ["src/models/"] }
    - { name: services, dirs: ["src/services/"] }
    - { name: routes, dirs: ["src/routes/"] }
  boundaries:
    - { from: models, to: [routes] }

heal:
  enabled: true
  healers:
    lint: { auto_fix: true }

integrations:
  claude_code: { enabled: true }

Implementation Roadmap

Version	Features	Status
v0.1.x	Core framework, 3 quality checks (lint, type, test), budget tracking	Shipped
v0.2.0	Budget 2.5x increases, multi-provider routing, semantic cache, calibration	Shipped
v0.2.1	5 new quality checks, weighted scoring, baseline regression deltas	Shipped
v0.2.2	Architecture diagrams, code deduplication, type error fixes, function refactoring	Shipped
v0.3.0	Cognitive complexity, mutation testing, flaky test detection	Planned
v0.4.0	Change failure rate, agent edit accuracy, cross-project dashboards	Planned
v1.0.0	Stable API, full TypeScript parity, plugin architecture	Planned

IDE Integrations

armature hooks --claude-code      # .claude/settings.local.json
armature hooks --cursor           # .cursor/rules
armature hooks --copilot          # .github/copilot-instructions.md
armature hooks --github-actions   # .github/workflows/armature.yml
armature hooks --pre-commit       # .pre-commit-config.yaml

The Harness Engineering Model

Armature implements the harness engineering 2x2 grid:

	Computational (fast, deterministic)	Inferential (LLM-based)
Feedforward (guides)	armature.yaml rules, architecture config, type hints	CLAUDE.md rules, spec constraints
Feedback (sensors)	ruff, mypy, boundary checks, conformance, GC sweeps	LLM code review, eval judges

Claude Code Skills

Armature provides slash commands for Claude Code:

• /armature-check -- Run quality sensors
• /armature-heal -- Self-healing pipeline
• /armature-gc -- Garbage collection sweep
• /armature-budget -- Cost tracking and reporting

Budget Control

Track and optimize AI coding costs:

# Log usage
armature budget --spec SPEC-001 --phase build --tokens 50000 --cost 1.25

# Generate report
armature budget --report SPEC-001

Armature analyzes phase distribution, per-request token usage, and suggests optimizations: batch file reads, narrow context, progressive disclosure.

References

• Ossature -- Spec-driven development with validate/audit/build
• OpenAI Harness Engineering -- Harness patterns for AI coding agents
• Martin Fowler's Exploring Gen AI -- Bockeler/Fowler harness engineering series

License

MIT